<p><strong>TLDR:</strong> Anthropic launched computer use in Claude Cowork and Claude Code, letting Claude control your mouse, keyboard, and screen on macOS as a research preview for Pro and Max subscribers. Stripe’s internal Minions system is generating over 1,300 pull requests per week, all human-reviewed. OpenCode got banned by Anthropic for faking auth headers, while T3 Code took the compliant route by wrapping the official CLI. Gemini 3.1 Pro landed in GitHub Copilot across JetBrains, Xcode, and Eclipse.</p>
<hr>
<h2>Claude gets mouse and keyboard control</h2>
<p>Anthropic shipped computer use in Claude Cowork and Claude Code as a research preview for Pro and Max subscribers on macOS. Claude tries connected integrations first — Slack, Calendar, and similar tools — and only asks to take over your screen when no connector exists. You can also set recurring tasks: scan email every morning, pull a report every Friday. What’s notable is how much of what previously required custom agentic frameworks — remote task dispatch, scheduled jobs, persistent memory — is now just built into the platform. The guardrails are real though: prompt injection detection, per-app permission requests, and Anthropic’s own recommendation to keep sensitive data away from it for now.</p>
<h2>Stripe’s Minions are generating 1,300+ PRs per week</h2>
<p>Stripe built an internal autonomous coding agent system called Minions that handles end-to-end software tasks from a single prompt. Every PR is human-reviewed before merging, but the volume is striking. This is one of the clearest public data points yet on what agentic coding looks like at scale inside a production engineering org.</p>
<h2>OpenCode banned, T3 Code takes the compliant path</h2>
<p>Anthropic banned OpenCode for faking headers and managing OAuth flows independently — essentially impersonating the official client to access Claude subscriptions. T3 Code went the other direction: it wraps the local Claude Code CLI without implementing its own auth, which keeps users compliant with Anthropic’s terms. The broader issue is that Anthropic’s policies around third-party integrations were vague and poorly communicated, leaving developers building on the SDK in an uncomfortable gray zone. The Claude Code $200/month plan reportedly subsidizes up to $5,000 in compute, which explains why Anthropic is watching this closely.</p>
<h2>Gemini 3.1 Pro lands in GitHub Copilot</h2>
<p>Gemini 3.1 Pro is now available in public preview for GitHub Copilot subscribers across <a href="http://github.com" target="_blank" rel="noopener nofollow">github.com</a>, GitHub Mobile, VS Code, Visual Studio, JetBrains IDEs, Xcode, and Eclipse. It’s accessible via the chat model picker in agent, ask, and edit modes. Copilot Business and Enterprise admins need to opt in through settings to enable it for their orgs.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Claude computer use + Dispatch:</strong> The /schedule command lets you create recurring cloud-based jobs for Claude directly from the terminal — Anthropic is using it internally for automation tasks.</li>
<li><strong>MCP tools on demand:</strong> Claude, Cowork, and Claude Code now load MCP tools on demand instead of stuffing all of them into context upfront, which should meaningfully reduce context bloat.</li>
<li><strong>VS Code mid-flight intervention:</strong> VSCode 1.110 lets you send corrective messages to Copilot agent mode while it’s still generating, plus queue multiple follow-up requests — no more passive waiting between prompts.</li>
<li><strong>GPT-5.4 vs Claude Opus 4.6:</strong> A benchmark comparison across 12 evals shows GPT-5.4 leading on terminal coding (75.1% vs 65.4% on Terminal-Bench 2.0) and computer use, while Opus 4.6 edges ahead on long-context workloads due to flat pricing. GPT-5.4 is 40-50% cheaper at base rates but doubles past 272K tokens.</li>
<li><strong>Lovable at $400M ARR:</strong> The vibe coding platform is actively hunting acquisitions and team acquisitions, growing at 200,000 new projects daily. They also shipped Security Checker 2.0 — four automated scanners running on every project before publish.</li>
<li><strong>Android Studio goes agentic:</strong> New features include a project assistant for scaffolding, UI check mode, and a “journeys” feature that generates code or tests from natural language user flow descriptions.</li>
<li><strong>cc-lens:</strong> Open source real-time monitoring dashboard for Claude Code — reads from ~/.claude/ with no cloud or telemetry, launched via npx cc-lens, covers token usage, costs, session replays, and activity heatmaps.</li>
<li><strong>Persona prompting hurts accuracy:</strong> A USC paper found that telling an LLM it’s an expert reduced MMLU accuracy from 71.6% to 68.0%. Skip the persona for factual and coding tasks; it only helps for alignment-dependent work like safety and structured output.</li>
<li><strong>GitHub Copilot in dotnet/runtime:</strong> After ten months and 878 CCA pull requests, the .NET team reports a 67.9% merge rate. Cleanup tasks hit 84.7% success; performance tasks bottom out at 54.5%. The bottleneck has shifted from code generation to code review.</li>
<li><strong>chardet relicensing controversy:</strong> The sole maintainer used Claude to rewrite the Python chardet library from scratch and relicensed it from LGPL to MIT. Bruce Perens concluded current copyright law can’t stop AI-mediated relicensing — a de facto loss for copyleft.</li>
<li><strong>Autoresearch with Claude Code:</strong> Applying Karpathy’s Autoresearch framework to an ML project, an agent ran 42 experiments in one day and reduced mean rank by 54%. The biggest win came from the agent spotting a bug in temperature parameter clamping.</li>
<li><strong>Black Duck Signal:</strong> New agentic security tool using AI agents trained on decades of human-validated security data to catch cross-file dataflow issues and business logic errors that rule-based AST tools miss. Integrates via MCP with IDEs and CI/CD pipelines.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

Anthropic launched computer use in Claude Cowork and Claude Code as a research preview for Pro/Max subscribers on macOS, enabling mouse/keyboard control with guardrails like prompt injection detection. Stripe's internal Minions system is generating 1,300+ human-reviewed PRs per week. OpenCode was banned by Anthropic for faking auth headers, while T3 Code took a compliant approach by wrapping the official CLI. Gemini 3.1 Pro is now available in GitHub Copilot across multiple IDEs. Additional highlights include MCP on-demand tool loading, VS Code mid-flight Copilot intervention, Lovable reaching $400M ARR, a USC study showing persona prompting hurts LLM accuracy, and the chardet library relicensing controversy via AI rewrite.

Claude gets computer use, Stripe's agents ship 1,300 PRs a week